Mapping Cyclist Safety in New York City

Is there a relationship between bike path access and cyclist injuries?

I was inspired by this CityLab article detailing the impending transportation crisis due to the MTA L Line shutdown. Specifically, I am interested in observing whether the shutdown contributes to an increase in cyclist injuries due to motor vehicle collisions as about 275,000 L Line riders seek alternatives.

NYC DOT has predicted about 2% of L Line riders to switch to cycling, while about 79% are predicted to move to other subway lines. I could investigate the dispersion of L Train riders to other options as a whole if data is available, but for now I will focus on cyclist safety as it’s more approachable with currently available data.

---

Let’s start by looking at some performance metrics for the MTA subway network as a whole.

Some of the performance metric definitions are opaque, so I’ll specify where necessary.

Terminal On-Time Performance measures the percentage of trains arriving at their destination terminals as scheduled. A train is counted as on-time if it arrives at its destination early, on time, or no more than five minutes late, and has not skipped any planned stops. TOTP is a legacy metric that provides a measure of trains arriving within the standard, and not a direct measure of customer travel time, particularly since relatively few customers travel all the way to the end of a line.

Wait Assessment measures how regularly the trains are spaced during peak hours. To meet the standard, the headway (time between trains) can be no greater than 25% more than the scheduled headway. This provides a percentage of trains passing the standard, but does not account for extra service operated, is not weighted to how many customers are waiting for the trains at different stations, does not distinguish between relatively minor gaps in service and major delays, and is not a true measurement of time customers spend waiting on the platform.

Mean Distance Between Failures (MDBF) reports how frequently car-related problems such as door failures, loss of motor power, or brake issues cause a delay of over five minutes. It is calculated by dividing the number of miles train cars run in service by the number of incidents due to car‐related problems.

---

Now let’s look at some two-wheeled data.

And now a map of cyclist injuries and deaths over the last few years.

Series: ts_injuries 
ARIMA(0,1,3)(0,1,1)[52] 

Coefficients:
          ma1     ma2      ma3     sma1
      -0.6902  0.0121  -0.2195  -0.5122
s.e.   0.0831  0.1001   0.0821   0.1416

sigma^2 estimated as 337.8:  log likelihood=-651.65
AIC=1313.3   AICc=1313.71   BIC=1328.31

Training set error measures:
                    ME     RMSE      MAE       MPE     MAPE      MASE
Training set -1.589831 15.57204 10.44011 -4.812162 13.84195 0.5500583
                    ACF1
Training set -0.01016123

---

Planned Updates / Notes:

This is an ongoing project that I am doing in my free time. If you have any constructive criticism please feel free to reach out, especially with suggestions about spatial regression methodology.

• Methodology
  • Need a way to do geospatial predictive modeling with a dataset that only has 1s, no 0s
  • How to understand relationship between accidents when we only have accident data, not non-accidents
  • Consider using Citibike data as a proxy to cyclist population, but do they behave differently?
  • Look into spatial prediction methods (spatial autocorrelation, spatial regression, etc.)
  • What trends are there in types of bike paths being constructed over time?
• Time-Series
  • More rigorous analysis of cyclist injuries
  • Use ARIMA model to remove seasonality, conclude whether injuries are increasing
  • Are injuries increasing because more people are biking or something else?
    
• Geospatial
  • Consider changing map to chloropleths (and use accidents per capita)
  • Find a way to visualize changes over time, whether through leaflet or gganimate or otherwise
  • Need function that can calculate a point’s (accident) nearest polyline (bike path) and return features (type of path)
    
• Visualizations
  • Change drag behavior (for mobile) when plotly is updated to 1.43